Shipping BAIO week after week, following AI × bio up close, makes me more optimistic about the future than I already was. There is so much ambition, ingenuity and sheer creativity in this field. This issue is no different.

☑️ Biohub trained a model to guess missing amino acids, and it appears to have learned an internal map of protein biology.

☑️ David Liu’s group took the enzyme that prime editors use to write new DNA letters and rebuilt large parts of it with AI - and it worked.

☑️ ASI created a benchmark around a question many of us have been asking lately: why the hell won’t Claude answer my ordinary biology questions?

☑️ Insilico and Human Longevity say they want to build a “super-intelligence” model for aging from longitudinal human data.

☑️ And by doing boring but hugely important infrastructure work, Perceptic wants to become the connective tissue that lets drug developers shine.

Let’s dive in.

CZ Biohub says it has built a world model for proteins. Calling something a “world model” is in vogue right now, so let’s unpack what they actually did here.

The protein language model, ESMC - short for Evolutionary Scale Model Cambrian - was trained on a simple task: predict missing amino acids in protein sequences. So far, so…predictable. But after training on about 2.8 billion sequences, “drawn from across all of life”, it appears to have learned a useful internal map of protein biology.

Not just which amino acids tend to appear together, but how proteins fold, what they do, which recurring tricks they use - catalytic sites, ligand-binding regions, stabilizing loops - where they sit in the cell, and how distant protein families connect across evolution.

The “world” of proteins, if you will.

Biohub reports that as training compute increased, the model’s internal representation of biology improved in a predictable way. “We identified a scaling law […] In ESMC, this powers linear returns with scale, leading to a new state of the art in protein representations,” the team writes.

ESMC is one part of a larger release. Biohub also built ESMFold2, a structure model that uses ESMC’s internal protein map to predict how proteins fold and how they interact with other molecules.

And with ESMFold2, Biohub reports another scaling trick. ESMFold2 can loop over a prediction multiple times at inference - basically spending more compute while it is solving the structure. That improved antibody-target predictions, one of the hard problems in protein modeling. In a preprint released with the launch, the authors report the same pattern in binder design: more computational search before the lab produced better experimental hit rates.

Using ESMFold2, the team picked five clinically relevant targets from cancer and immunology - immune checkpoints and cell-surface receptors, the kinds of proteins many antibody drugs already try to block or activate.

Instead of physically screening huge libraries, they searched computationally first. For each target, ESMFold2 helped rank tens of thousands to more than 100,000 possible binders. Then the team sent only 84 top designs per target and format into the lab.

That small lab test worked unusually well. For minibinders - tiny proteins designed from scratch - 36-88% attached to their targets. For antibody-like binders, a harder format because the binding loops are more flexible, 15-29% worked. Several grabbed their targets tightly enough to fall in the nanomolar or even picomolar range drug developers care about. One PD-L1 binder also did more than stick: in a cell assay, it restored T cell signaling in the same range as a version of atezolizumab, an approved cancer immunotherapy.

“What we've shown is that these models have learned such a high-fidelity world model of biology that you can design protein interfaces computationally, take them into the laboratory, and they function as predicted,” says Alex Rives, Biohub’s Head of Science, in a press release.

Another part Biohub stresses is speed. Hundreds of thousands of antibody candidates can be screened in a few hours of computational time. The lab-validated protein binders above were designed in days, not weeks or months.

The release also includes ESM Atlas, which is Biohub’s attempt to turn the same protein model into a searchable map of life’s proteins. It covers 6.8 billion protein sequences and 1.1 billion predicted structures - structures Biohub says were computed in about two weeks.

Why it matters: BAIO covered Biohub’s $500 million virtual-cell push in Issue 21. This is the protein side of the same bet: make more of biology computational before the lab gets involved. If the approach holds up, the biggest payoff may be where attention is scarce - rare diseases, molecularly specific cancers, and targets that have not attracted much traditional pharma investment.

Did you know? Biohub is making the models and the ESM Atlas available for researchers to use.

Prime editing is the elegant version of genome editing: find a DNA sequence and rewrite it, without cutting both DNA strands. The catch is that the editor itself is a big protein machine. It has to fold properly, stay stable, and be produced at useful levels inside cells.

In a Nature Biotechnology paper, David Liu’s group at the Broad Institute found that earlier prime editors improved through laboratory evolution - repeated rounds of mutation and selection in the lab - had a trade-off. They edited better, but some became less stable and showed lower protein levels. So the team used ProteinMPNN, an AI protein-design model (that BAIO readers have heard about several times before), to redesign reverse transcriptase - the enzyme that writes the new genetic letters - while leaving key regions alone.

The result is a new set of prime editors called PE8. The key point is that the AI redesign changed large parts of the reverse transcriptase while leaving the working core alone. That is risky: enzymes depend on precise shapes, and changing dozens or hundreds of amino acids can easily make them misfold or stop working.

Here, most did not break. In cultured cells, 165 of 174 redesigned enzymes cleared the team’s 5% screening threshold. About 30% worked better than the editors they started from. The best PE8 variants also produced more usable editor protein inside cells - up to twice as much after mRNA delivery.

That is important because many therapeutic prime-editing setups deliver the editor only transiently, often as mRNA. If the same dose makes more usable editor protein inside the cell, you get more editing from that short delivery window. Across 700 ClinVar disease-related edits, the best redesigned versions improved average editing. In mice, PE8 variants improved liver editing up to 2.9-fold over earlier state-of-the-art editors.

Why it matters: BAIO covered in Issue 4 how AI is becoming essential infrastructure for CRISPR. This is a concrete example: AI improved the editing machine itself, making prime editors more stable and more potent under delivery conditions that matter for therapies.

Did you know? The code used for the redesign is on GitHub, and the authors say the PE8 plasmids will be available through Addgene.

For a while now, I’ve been researching AI and biosecurity for an article, including testing how the latest models respond to questions around DNA-ordering safeguards and virus construction. Some answered more than they probably should. Some flagged my account. Others refused almost everything.

Anthropic’s models have increasingly felt like they belong in that last camp - and not only for obvious dual-use prompts. Since Claude Opus 4.7, I’ve run into refusals on ordinary biology questions too, which is not ideal when you write a newsletter about AI × bio.

Turns out I’m not alone. Applied Scientific Intelligence, or ASI - whose Alexandria literature agent BAIO covered in Issue 21 - has released RefusalBench, testing how 19 frontier models handle biological research prompts.

Across 141 biological research prompts, the task stays roughly the same, but the risk level changes: benign, borderline, or dual-use. In other words, the benchmark asks whether models can tell the difference between ordinary computational biology and work that starts moving toward misuse.

Strict refusal rates ranged from 0.1% to 94.6% on the same benchmark. Anthropic stood out on the restrictive end: the paper says their models were roughly 21 times more likely to refuse than the non-Anthropic baseline. ASI’s public write-up also highlights benign protein-design prompts where the most restrictive models refused at very high rates, even though the prompts used human therapeutic targets and no biosecurity flags.

As you can probably guess, refusing more is not the same as being safer. Some models were too permissive: three failed to refuse prompts the authors say clearly should have been rejected. Others showed what the paper calls a “hedge-but-help” pattern: they add safety language, then still provide useful information. Nine of 18 frontier models showed that pattern on dual-use prompts.

Why it matters: As ASI notes these systems are increasingly becoming part of how research gets done. Which means refusal behavior becomes part of the infrastructure too. Bad calibration cuts both ways: over-refusal slows legitimate science, under-refusal creates real biosecurity risk.

Did you know? RefusalBench is open on GitHub, and ASI says labs can run it against their own orchestration models to see where they over-refuse or under-refuse across biology risk tiers.

Insilico Medicine and Human Longevity are teaming up to build what they call a foundation model for human longevity. The deal runs through Human Life Foundation Models, a new company established by Human Longevity, and is described as a multi-million-dollar co-development collaboration.

The pitch is simple, and very ambitious: combine Insilico’s AI drug-discovery machinery with Human Longevity’s de-identified multi-omic, imaging, and longitudinal health datasets from thousands of people. The goal is to build models that can spot disease risk earlier, model aging biology, and point toward new longevity therapeutics.

Insilico, a recurring character in this newsletter, is mostly known as an AI drug developer company. But its long term goal has always been to fight aging.

It’s worth pointing out that this is a press release, not a paper. No model, benchmark, or validation data were released. And the press-release language is very large - “super-intelligence,” “predict complex disease decades before it occurs,” and so on.

Why it matters: Longevity has always had a data problem, something we discussed in this newsletter just a few short weeks ago. In Issue 25, Morgan Levine argued that AI × longevity needs richer, more diverse aging data, while Martin Borch Jensen argued that the data has to sit at the right biological layer for the question being asked. This collaboration is a direct bet on that idea. Aging is slow, systemic, and can’t be modeled from snapshots. If Human Longevity’s longitudinal omics, imaging, and clinical data is strong enough, and if Insilico can turn it into models that predict something useful, this could become real infrastructure for longevity research.

Did you know? Insilico is hiring.

Perceptic came out of stealth with $12 million to build what it calls an AI operating system for drug development. The company was founded by Tilman Flock, Martin Copes, and Zaki Trache, former Palantir executives who worked on its AI platform and life-sciences business.

Perceptic wants to sit across the messy middle of pharma: asset scouting, scientific due diligence, clinical data extraction, trial history, competitive intelligence, and internal decision-making. The company pitches itself as the connective tissue between point AI tools and the data pharma companies actually use to make decisions.

Perceptic’s software is reportedly already being used by multiple top-tier pharma companies, including CSL.

Why it matters: Not everything in drug discovery has to do with biology. And so Perceptic is not trying to replace the lab. It is trying to make the decision layer around the lab faster.

Did you know? Perceptic is hiring.

Goodfire’s EVEE, which BAIO covered in Issue 17, is now available as an MCP server for AI agents. It lets Claude and other agents query pathogenicity predictions, disruption profiles, and mechanistic explanations for 4.2 million ClinVar variants. Use it to look up a variant, compare up to ten at once, or ask what biological annotations the model thinks are disrupted. Install from GitHub.

☑️ YC-backed Panacea Bio launches AI-native FDA regulatory services.

☑️ ClinSeekAgent gives medical agents active evidence-seeking tools.

☑️ Obvious Ventures publishes a “Biological Awareness” thesis.

☑️ Terminal-Bench Science asks scientists to contribute real workflows for testing AI agents.

☑️ Arc’s PerturbSpace adds spatial context to CRISPR screens.

☑️ Mo Lotfollahi frames CRISPR screening as a lab-in-the-loop design problem.